AI-supported design of effective intervention strategies

Important results from the project

The project has developed and implemented a reinforcement learning agent in a simulation model fitted against the number of reported cases of covid-19 in 2020. The agent´s ability to identify effective non-medical intervention strategies to reduce the spread of infection has been studied to find intervention strategies that are evaluated against more objectives than only the number of infected or deceased. In particular, the balance between physical distancing, mobility in society, and the spread of infection has been taken into account.

Expected long term effects

We developed a framework for simulation and reinforcement learning during the project, which can be used in future simulation projects at the agency. The framework found efficient and well-balanced intervention strategies that take multiple decision-criteria into account, such as the societal disease burden and negative effects of physical distancing. The framework has great potential to be used as a decision-support tool during a pandemic by dynamically supporting decision-makers in selecting an intervention strategy.

Approach and implementation

The project consisted of three work packages, i) choice of reinforcement learning method, ii) implementation, and iii) evaluation. A series of workshops were conducted; however, most were distance-based due to the covid-19 pandemic. We investigated available methods, limitations in the available methods, and program libraries for reinforcement learning during the project. The reinforcement agent was implemented, and the model was trained and evaluated against existing benchmarks.